Meta-Learning for Stacked Classification

In this paper we describe new experiments with the ensemble learning method Stacking. The cen-tral question in these experiments was whether meta-learning methods can be used to accurately predict various aspects of Stacking’s behaviour. The resulting contributions of this paper are two-fold: When learning to predict the accuracy of stacked classifiers, we found that the single most important feature is the accuracy of the best base classifier. A simple linear model involving just this feature turns out to be surprisingly accu-rate. The associated regression line has a gra-dient larger than one, hinting that, in the limit, Stacking is indeed better than the best included base classifier. When learning to predict signifi-cant differences between Stacking and three other ensemble learning methods, we have found sim-ple models, all but one of which are based on sin-gle features which can be efficiently computed directly from the dataset. These models can be used to decide in advance which ensemble learn-ing method to use on a given dataset, since nei-ther of them is always the best choice. 1